KR20190135916A - Apparatus and method for determining user stress using speech signal - Google Patents

Abstract

The present invention provides an apparatus and a method for determining user stress using a speech signal which can accurately determine stress of a user. The apparatus for determining user stress using a speech signal comprises: a power spectrum conversion unit to convert a speech signal divided into a plurality of frames into a power spectrum; a filter bank unit including a plurality of Mel-filter banks each having a designated frequency band and pattern to filter the power spectrum of the plurality of frames to acquire a plurality of Mel-filter bank energies; a feature vector acquisition unit to apply a plurality of weights determined in accordance with a pre-learned pattern estimation method to corresponding Mel-filter bank energies among the plurality of Mel-filter bank energies to acquire feature vectors for each frame; a frame feature extraction unit to sequentially encode a plurality of feature vectors for each frame in accordance with a pre-learned pattern estimation method to extract frame features, and encoding the feature vectors for each frame with previously acquired frame features to acquire frame features; and a speech feature extraction unit to receive a final frame feature among the frame features acquired by the frame feature extraction unit, and extracting a speech feature corresponding to stress of a user from the final frame feature in accordance with a pre-learned pattern estimation method.

Description

Apparatus and method for determining user stress using speech signal}

The present invention relates to a user stress determination apparatus and method, and to a user stress determination apparatus and method that can determine the presence or absence of the stress of the user using a voice signal on the deep learning.

Generally, a technique for determining stress using a speech signal is performed through a statistical method of extracting a feature related to emotion or stress recognition from a speech signal and modeling a relationship between the extracted feature and a predetermined stress label.

At this time, feature extraction is mainly performed based on a method proposed by a skilled expert. However, in this case, although there is an advantage in that the utility is large, there is a limitation that it is not sure whether the selected feature extraction method shows the best result for the data to be trained.

In addition, conventional statistical methods include a Gaussian Mixture Model (GMM) and a Support Vector Machine (SVM) algorithm. The GMM algorithm works by classifying which group a given data belongs to when given data in a test situation by representing the probability distribution of the data to be classified as the sum of a number of normal distributions. In the case of the SVM algorithm, when two groups of data are present, the data is classified by optimizing a hyperplane between the two.

However, these statistical methods have a limitation in that it is difficult to model changes in the time axis, which makes it difficult to discriminate stress from speech signals whose characteristics change very rapidly in the time domain.

Korean Unexamined Patent No. 10-2017-0117019 (published Oct. 20, 2017)

It is an object of the present invention to provide an apparatus and method for determining a user stress using a speech signal that can accurately determine a user's stress by extracting an optimized feature vector from the speech signal.

In accordance with an aspect of the present invention, a user stress determination apparatus using a voice signal includes: a power spectrum converter configured to convert a voice signal divided into a plurality of frames into a power spectrum; A filter bank unit including a plurality of mel-filter banks each having a predetermined frequency band and a pattern, respectively, filtering the power spectrum of each of the plurality of frames to obtain a plurality of mel-filter bank energies; A feature vector obtaining unit obtaining a feature vector for each frame by applying a corresponding mel-filter bank energy among the plurality of mel-filter bank energies to a plurality of weights determined according to a previously learned pattern estimation scheme; A frame feature extractor configured to sequentially encode a plurality of feature vectors for each frame according to a pre-learned pattern estimation method to extract frame features, and to obtain frame features by encoding previously obtained frame features together; And a speech feature extractor configured to receive a final frame feature among the frame features obtained by the frame feature extractor and to extract a voice feature corresponding to the user's stress from the final frame feature according to a pre-learned pattern estimation scheme. It includes.

The frame feature extracting unit is a pre-learned artificial neural network including a plurality of encoders that extract a current frame feature by receiving a corresponding frame feature vector and a previously obtained frame feature among a plurality of frame feature vectors. Can be implemented.

The speech feature extractor may combine a plurality of frame features extracted from each of the plurality of encoders and extract the speech feature from the plurality of frame features combined according to a pre-learned pattern estimation scheme.

The user stress determination apparatus obtains learning data labeled with a stress level during learning, extracts an error by comparing a voice feature of the acquired learning data with a labeled stress level, and extracts the obtained error from the frame feature extracting unit. And learning by back propagating to the speech feature extractor, and if the error between the learned speech feature and the speech feature obtained using the speech feature extractor and the labeled stress level exceeds a predetermined reference error, A learning unit trains the frame feature extractor and the speech feature extractor after updating the weight by back propagating an error with a vector obtainer; It may further include.

Each of the plurality of mel-filter banks may be implemented as a triangular filter having a different bandwidth from a mel-frequency scale predetermined for each frequency band.

According to another aspect of the present invention, there is provided a method for determining a user stress using a voice signal, comprising: dividing the voice signal into a plurality of frames and converting each of the plurality of frames into a power spectrum; Using a filter bank having a plurality of filters each having a predetermined frequency band and a pattern, filtering the power spectrum divided by frames to obtain mel-filter bank energy in each of the plurality of frames; Obtaining a feature vector by applying a weight determined for each frame by a pre-learned pattern estimation scheme to a plurality of mel-filter bank energies; Extracting a frame feature by sequentially encoding each of a plurality of mel-filter bank energies weighted in the feature vector according to a pre-learned pattern estimation method, and encoding a previously obtained frame feature together to obtain a frame feature; And receiving a final frame feature among the acquired frame features and extracting a voice feature corresponding to the user's stress from the final frame feature according to a pre-learned pattern estimation scheme. It includes.

Therefore, the apparatus and method for determining a user stress using a speech signal according to an embodiment of the present invention extracts an optimized feature vector from the speech signal and uses the artificial neural network to reflect previous information from the extracted feature vector. It can greatly increase the accuracy. You can also accurately model the parameters associated with stress in speech.

1 shows a schematic structure of a user stress determination apparatus using a voice signal according to an embodiment of the present invention.
2 and 3 are diagrams for describing an operation of the feature vector extracting unit of FIG. 2.
4 shows a detailed configuration of the feature extraction unit of FIG. 1.
5 illustrates a user stress determination method using a voice signal according to an embodiment of the present invention.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings which illustrate preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In addition, in order to clearly describe the present invention, parts irrelevant to the description are omitted, and like reference numerals in the drawings indicate the same members.

Throughout the specification, when a part is said to "include" a certain component, it means that it may further include other components, without excluding other components unless otherwise stated. In addition, the terms "... unit", "... unit", "module", "block", etc. described in the specification mean a unit that processes at least one function or operation, which means hardware, software, or hardware. And software.

1 shows a schematic structure of a user stress determination apparatus using a voice signal according to an embodiment of the present invention, Figures 2 and 3 are views for explaining the operation of the feature vector extraction unit of Figure 2, Figure 4 is a view The detailed structure of the feature extraction part of 1 is shown.

Referring to FIG. 1, the apparatus for determining a user stress using a voice signal according to the present exemplary embodiment includes a voice signal acquirer 110, a feature vector acquirer 120, and a stress determiner 130.

The voice signal acquisition unit 110 acquires a voice signal of a user who is a stress determination target. The voice signal acquisition unit 110 may be implemented by a device such as a microphone to obtain a voice signal, but may be implemented by various devices such as a communication device for obtaining a voice signal through wired / wireless communication or a storage device in which the voice signal is stored in advance. Can be.

The feature vector acquirer 120 obtains the feature vector in a predetermined manner by receiving the voice signal acquired by the voice signal acquirer 110. The feature vector acquirer 120 divides the voice signal of the time domain into frame units having a predetermined length (for example, 5 ms), and analyzes energy for each frequency band for each divided frame to obtain a feature vector for the voice signal. Acquire.

For example, the feature vector acquirer 120 may obtain a feature vector by extracting mel-filter bank energy for each frame based on a mel-frequency cepstral coefficients (MFCC) technique from a speech signal.

Mel-frequency cepstral coefficients (MFCC) is a typical technique for acquiring a feature vector from a speech signal. A number of preset preset frequency bands (Mel-frequency scale) have a corresponding size in consideration of human hearing characteristics. Mel-filter banks are used to divide the speech signal into several frequency bands, obtain the energy of the filtered signal in each mel-filter bank, and post-process the obtained energy with various statistical techniques. The technique of obtaining a vector.

The MFCC technique is an extraction method that allows a speech signal to have a different meaning according to a frequency range filtered by a plurality of mel-filter banks, and is used in various fields such as speech recognition and speech synthesis.

However, in the present exemplary embodiment, the feature vector acquirer 120 uses the energy per frequency band itself filtered by the plurality of mel-filter banks as a feature vector. In other words, no post-processing is performed according to statistical techniques. This is to allow the stress determination unit 130 to be described later to determine the stress by extracting the features of the original voice signal itself.

However, in the present embodiment, the feature vector acquirer 120 obtains the feature vector by applying a weight w corresponding to the energy of each frequency band divided by frame so as to accurately determine the stress of the user. In this case, a plurality of weights w for each energy of each frequency band may be updated and optimized during the learning process of the stress determination device.

The feature vector acquirer 120 may include a frame divider 121, a power spectrum converter 123, a filter bank 125, and a weight applier 127.

The frame separator 121 receives the voice signal through the voice signal acquisition unit 110 and divides the frame into a plurality of frames having a predetermined time unit (for example, 5 ms).

The power spectrum converter 123 converts a voice signal divided into a plurality of frames into a power spectrum. 2A illustrates an example of a power spectrum converted for each frame. For example, the power spectrum converter 123 may obtain a power spectrum by performing a fast Fourier transform (FFT) on each of a plurality of frames.

The filter bank unit 125 includes a plurality of mel-filter banks, each configured to filter a specified frequency band, and each of the plurality of mel-filter banks filters a power spectrum to display a mel-indicating energy for each frame and frequency band. Obtain filter bank energy. Here, the plurality of mel-filter banks may be implemented as triangular filters having different bandwidths on a mel-frequency scale predetermined for each frequency band, as shown in (b) of FIG. 2. A plurality of mel-filter banks may perform filtering for each frame of L to extract mel-filter bank energy for each frame and for each frequency band, as shown in FIG.

The filtering function of each of the plurality of mel-filter banks may be represented by H _m (k), and the power spectrum of each frame may be filtered according to Equation 1.

Where k denotes an index of the power spectrum, m denotes a mel-filter bank index, and f (·) denotes a frequency band of the mel-filter bank index m.

And Mel-filter bank energy (e _m) is a number of enamel each frame is obtained by the filter bank to filter the sum of the power spectrum in each as in the formula (2).

Where s (k) represents the power spectrum of the kth index in any frame.

That is, the mel-bank energy (e _m ) obtained through the filtering by the m-th mel-filter in any frame is calculated according to Equation 2. This allows the mel-filter bank energy (e _m ) to be calculated in one frame.

The weight applying unit 127 is implemented as a pre-learned artificial neural network and applies a weight w _m corresponding to the plurality of mel-filter bank energies e _m obtained from the filter bank unit 125. In this case, the weight applying unit 127 may apply different weights w _{m for} each mel-filter bank according to a previously learned pattern. In addition, the weight applying unit 127 has a mel-filter bank energy w _m * to which a weight w _m is applied for each frame. e _m ) to obtain a feature vector consisting of *.

In this case, the weight w _m may be applied to be weighted to a plurality of frequency bands of energy filtered in the mel-filter bank, but may be applied as a gain of the plurality of mel-filters of the mel-filter bank. That is, by applying a weight (w _m ) directly to each of the plurality of mel-filter banks, the energy band-specific energy band (e _m * w _m ) to which the weight (w _m ) is applied is obtained, and the acquired frequency-specific mel-filter banks It can be configured to sum energy.

3 is a view illustrating a concept in which the feature vector obtainer 120 acquires a feature vector. In FIG. 3, (a) is divided by the frame separator 121 and converted by the power spectrum converter 123. The multiple power spectra s (k) for the frame are shown. (b) shows the filtering function H _m (k) of each of the plurality of mel-filter banks, and (c) shows that the plurality of power spectra s (k) is filtered by each of the plurality of mel-filter banks. Mel-filter bank energy (e _m ) is shown, and (d) represents a process of applying a weight (w _m ) to each mel-filter bank energy (e _m ). Finally, (e) is a feature vector (v _f ) consisting of weighted mel-filter bank energies (we _m = e _m * w _m ) with weights (w _m ) corresponding to each of a plurality of mel-filter banks in each frame. ). As shown in FIG. 3, when there are M mel-filter banks, the frame-specific feature vector v _f includes M weighted mel-filter bank energies we ₁ to we _M.

Meanwhile, the stress determination unit 130 is implemented with an artificial neural network in which a pattern estimation method is pre-learned, and receives a plurality of frame-specific feature vectors v ₁ to v _F obtained from a plurality of frames, respectively, and applies a plurality of frames. Stress related features f ₁ to f _F are extracted from the feature vectors v ₁ to v _F to determine the stress of the user. For example, the stress determination unit 130 may be implemented as a recurrent neural network (RNN) or a long short term memory (LSTM), which is an artificial neural network that reflects a previously extracted feature of an artificial neural network when the current feature is extracted.

Referring to FIG. 4, the stress determiner 130 includes a frame feature extractor 131 and a voice feature extractor 133.

The frame feature extractor 131 is implemented as an artificial neural network including a plurality of encoders EN ₁ to EN _F to encode corresponding feature vectors for each frame among the plurality of frame feature vectors v ₁ to v _F , respectively. Extract the frame features f ₁ to f _F. In FIG. 4, the frame feature extractor 131 is implemented as an LSTM. In this case, the plurality of encoders EN ₁ to EN _F may be viewed as LSTM cells.

Each of the plurality of encoders EN ₁ to EN _F passes the extracted frame features f ₁ to f _F to the next stage of encoders EN ₁ to EN _F , and the last stage of encoder _F EN The final frame feature f _F is transmitted to the speech feature extractor 133.

Wherein a plurality of Delivering the encoder is extracted frame features _{(EN 1 ~ EN F-1} ) (f 1 ~ f F-1) to the encoder at the next stage the previously extracted frame features _{(f 1 ~ f F-1} ) This is to be considered together when extracting the next frame features f ₂ to f _F. This is to allow time information to be reflected by correlating features between a plurality of frames divided by a predetermined time unit. Therefore, it can be seen that the characteristics of the previous frame features f ₁ to f _F-1 are accumulated and reflected in the final frame feature f _F.

The speech feature extractor 133 is also implemented as a pre-learned artificial neural network to extract the speech feature s corresponding to the user stress from the final frame feature f _F.

In some cases, the speech feature extractor 133 extracts the speech feature s by concatenating all frame features f ₁ to f _F extracted by the plurality of encoders EN ₁ to EN _F. It may be configured to.

In addition, when the plurality of encoders EN ₁ to EN _{F of the} frame feature extractor 131 extract the frame features f ₁ to f _F , the plurality of encoders EN ₁ to EN _F further extract the hidden features h ₁ to h _F. The extracted hidden features may be transferred to the encoder of the next stage so that the previously extracted hidden features h ₁ to h _F may be considered together when the next frame features f ₁ to f _F are extracted.

The voice feature extractor 133 may extract the voice feature s corresponding to the user's stress with or without a stress, or may extract the voice feature extractor with a stress level divided by a predetermined unit.

In the stress determination apparatus according to the present embodiment, the stress determination unit 130 and the weight applying unit 127 implemented as an artificial neural network must be previously learned and optimized as described above. The stress determination device may further include a learning unit 140 for learning of the stress determination unit 130 and the weight applying unit 127.

The learner 140 may acquire a plurality of learning data that is a voice signal having a stress level or a stress level, and provide the acquired learning data to the voice signal obtaining unit 110 during a learning process of the stress determination device. In addition, the stress determination result is received from the stress determination unit 130 to determine the error by comparing the determination result with the stress level or the stress level labeled in the training data, and the propagated error to the stress determination unit 130 by back propagating the stress The determination unit 130 may be trained. Hereinafter, a description will be made on the assumption that the stress level is labeled in the training data for convenience of explanation, but the learning data labeled as stress may be used as described above.

The learner 140 repeatedly learns the stress determiner 130 using a plurality of pieces of learning data, and at this time, a plurality of weights w ₁ corresponding to each of the plurality of mel-filter banks of the feature vector acquirer 120. ~ w _M ) is fixed to a previously obtained value for each frame. In addition, the learner 140 reverse-propagates the error according to the stress determination result for the plurality of learning data based on the current weights w ₁ to w _M to the stress determination unit 130 to perform the stress determination unit 130. To learn. Based on the learned stress determination unit 130, an error according to the stress determination result of the plurality of learning data is obtained again, and the obtained error is transmitted to the weight application unit 127, and a plurality of weights (w ₁ ˜) are obtained. w _M ) is updated.

When a plurality of weights w ₁ to w _M corresponding to the plurality of mel-filter banks are updated by the weight updating unit 150, the plurality of learning data are used again according to the updated weights w ₁ to w _M. To repeat the process of learning the stress determination unit 130. That is, whenever the weights w ₁ to w _M are updated, the stress determination unit 130 is trained again.

The learning unit 140 ends the learning of the stress determination device when the number of learning reaches a predetermined reference number or the error is within the predetermined reference error.

5 illustrates a user stress determination method using a voice signal according to an embodiment of the present invention.

Referring to FIGS. 1 to 4, when a method of determining a user stress using a voice signal according to the present embodiment is described, first, a plurality of learning data for learning a stress determination apparatus are obtained (S11). Here, the training data is a speech signal labeled with stress or stress level.

Then, the obtained training data is input to an untrained stress determination device to obtain a stress determination result (S12). When the stress determination result for the training data is obtained, an error is determined by comparing the obtained stress determination result with the stress level labeled in the training data, and the propagated error is propagated back to the stress determination unit 130 of the stress determination device to stress the stress. Train the discriminating unit (S13).

In operation S14, it is determined whether the stress determination unit 130 is trained using the obtained total learning data. If it is determined that learning on the entire learning data has not been performed, another learning data is input to the stress determining device again to obtain a stress determining result (S12). However, if it is determined that learning has been performed on all acquired learning data, it is determined whether the determined error is less than or equal to the predetermined reference error (S15).

If the error exceeds the reference error, the error is propagated back to the weight applying unit 127 of the feature vector obtaining unit 120 to update a plurality of weights w ₁ to w _M (S16). In addition, the stress determination unit 130 is trained using the entire learning data.

On the other hand, if the error is within the reference error, the learning of the stress determination device is terminated, and a user's voice signal to determine the stress is obtained (S17). The obtained voice signal is divided into a plurality of frames of a predetermined time unit (S18). Thereafter, each of the plurality of frames is converted into a power spectrum (S19).

Meanwhile, a plurality of mel-filter bank energies (e _m) are filtered by filtering a plurality of power spectrums converted for each frame using a plurality of mel-filter banks of the feature vector acquirer 120 according to a predetermined mel-frequency scale and a frequency bandwidth. ) Is obtained (S20). In addition, a plurality of weighted mel-filter bank energies (we _m ) are obtained by applying the weights (w _m ) obtained through learning to the obtained plurality of mel-filter bank energies (e _m ), and the obtained plurality of weighted mels are obtained. The frame-specific feature vectors v ₁ to v _F composed of the filter bank energy we _m are obtained (S21).

When the frame-specific feature vectors v ₁ to v _F are obtained, the frame features f ₁ to f are obtained from the frame-specific feature vectors v ₁ to v _F obtained for each of a plurality of frames according to a previously learned pattern recognition scheme. _F ) is extracted (S22). In this case, the plurality of frame features f ₁ to f _F may be extracted by reflecting the previously extracted frame features f ₁ to f _F-1 . When a plurality of frame features f ₁ to f _F-1 are extracted, the voice features are pre-learned from a final frame feature f _F among the extracted plurality of frame features f ₁ to f _F. (S) is extracted to determine the stress of the user (S23).

As a result, the apparatus and method for determining a user stress using a speech signal according to the present embodiment are characterized by a frame-specific feature vector (v) consisting of a plurality of weighted mel-filter bank energies (we _m ), each of which is weighted in a plurality of frames separated from the speech signal. ₁ ~ v _f) extract and a plurality of frame-by-frame feature vector (v ₁ ~ a v _f) prior extracted frame feature for _{(f 1 ~ f f-1} ) is the considered frame features (f ₁ ~ f _f) Extract the voice feature (s) corresponding to the stress of the user. Therefore, it is possible to accurately determine the stress of the user.

The method according to the invention can be implemented as a computer program stored in a medium for execution in a computer. The computer readable media herein can be any available media that can be accessed by a computer and can also include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, and includes ROM (readable) Dedicated memory), RAM (random access memory), CD (compact disk) -ROM, DVD (digital video disk) -ROM, magnetic tape, floppy disk, optical data storage, and the like.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible.

Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

110: voice signal acquisition unit 120: feature vector acquisition unit
130: stress determination unit 140: learning unit
121: frame separator 123: power spectrum converter
125: filter bank unit 127: weight applying unit
131: frame feature extractor 133: voice feature extractor

Claims (9)

A power spectrum converter for converting a voice signal divided into a plurality of frames into a power spectrum;
A filter bank unit including a plurality of mel-filter banks each having a predetermined frequency band and a pattern, respectively, filtering the power spectrum of each of the plurality of frames to obtain a plurality of mel-filter bank energies;
A feature vector obtaining unit obtaining a feature vector for each frame by applying a corresponding mel-filter bank energy among the plurality of mel-filter bank energies to a plurality of weights determined according to a previously learned pattern estimation scheme;
A frame feature extractor configured to sequentially encode a plurality of feature vectors for each frame according to a pre-learned pattern estimation method to extract frame features, and to obtain frame features by encoding previously obtained frame features; And
A voice feature extractor configured to receive a final frame feature among the frame features obtained by the frame feature extractor and to extract a voice feature corresponding to a user's stress from the final frame feature according to a pre-learned pattern estimation scheme; Device for determining the user stress comprising a. The method of claim 1, wherein the frame feature extraction unit
And a plurality of encoders configured to extract a current frame feature by receiving a corresponding frame feature vector and a previously obtained frame feature among a plurality of frame feature vectors. The speech feature extractor of claim 1, wherein the speech feature extractor comprises:
And combining the plurality of frame features extracted by each of the plurality of encoders, and extracting the speech feature from the plurality of frame features combined according to a pre-learned pattern estimation scheme. According to claim 1, wherein the user stress determination device
Acquire learning data labeled with a stress level at the time of learning, extract an error by comparing a voice feature of the acquired learning data with a stress level labeled with the learning data, and obtain the obtained error from the frame feature extractor and the Backpropagation to the speech feature extractor to learn,
When the error between the learned speech feature and the speech feature extracted using the frame feature extractor and the speech feature extractor exceeds a predetermined reference error, the error is propagated back to the feature vector obtainer. A learner configured to learn the frame feature extractor and the speech feature extractor after updating the weights; Device for determining the user stress further comprising. The method of claim 1, wherein each of the plurality of mel-filter banks are
A user stress determination device implemented by a triangular filter having a different Mel-frequency scale and a different bandwidth for each frequency band. Dividing the speech signal into a plurality of frames and converting each of the plurality of frames into a power spectrum;
Using the plurality of mel-filter banks each having a predetermined frequency band and pattern, respectively filtering the power spectrum of each of the plurality of frames to obtain a plurality of mel-filter bank energies;
Obtaining a feature vector for each frame by applying a corresponding mel-filter bank energy among the plurality of mel-filter bank energies to a plurality of weights determined according to a pre-learned pattern estimation scheme;
Extracting a frame feature by sequentially encoding a plurality of frame feature vectors according to a pre-learned pattern estimation method, and encoding a previously obtained frame feature together to obtain a frame feature; And
Receiving a final frame feature among the acquired frame features, and extracting a voice feature corresponding to the user's stress from the final frame feature according to a pre-learned pattern estimation scheme; User stress determination method comprising a. 7. The method of claim 6, wherein extracting the speech feature
Applying and combining all of the plurality of frame features including the last frame feature; And
Extracting the speech feature from the combined plurality of frame features according to a pre-learned pattern estimation scheme; User stress determination method comprising a. The method of claim 6, wherein the user stress determination method is
Learning phase; More,
The learning step
Obtaining training data labeled with a stress level;
To extract the error by comparing the speech feature of the acquired training data with the stress level labeled in the training data, and to extract the frame feature by back propagating the obtained error and to extract the speech feature. Updating the pattern estimation scheme;
If the error between the speech feature acquired based on the updated pattern estimation method and the stress level labeled in the training data exceeds a predetermined reference error, the frame feature is updated again by updating the weight by back propagating the acquired error. Updating a pattern estimation method for extracting a signal and a pattern estimation method for extracting the speech feature; User stress determination method further comprising. 7. The method of claim 6, wherein obtaining the mel-filter bank energy is
A method of determining a user stress, wherein the mel-filter bank energy is obtained by using a plurality of mel-filter banks each formed of a triangular filter having a different mel-frequency scale and a different bandwidth.

Images